Preserve ASAN-related symbols during LTO.

compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

@@ -1058,6 +1058,13 @@ extern "C" void LLVMRustPrintPasses() {
 extern "C" void LLVMRustRunRestrictionPass(LLVMModuleRef M, char **Symbols,
                                            size_t Len) {
   auto PreserveFunctions = [=](const GlobalValue &GV) {
+    // Preserve LLVM-injected, ASAN-related symbols.
+    // See also https://github.com/rust-lang/rust/issues/113404.
+    if (GV.getName() == "___asan_globals_registered") {
+      return true;
+    }
+
+    // Preserve symbols exported from Rust modules.
     for (size_t I = 0; I < Len; I++) {
       if (GV.getName() == Symbols[I]) {
         return true;

tests/codegen/sanitizer/address-sanitizer-globals-tracking.rs

@@ -0,0 +1,43 @@
+// Verifies that AddressSanitizer symbols show up as expected in LLVM IR with `-Zsanitizer`.
+// This is a regression test for https://github.com/rust-lang/rust/issues/113404
+//
+// Notes about the `compile-flags` below:
+//
+// * The original issue only reproed with LTO - this is why this angle has
+//   extra test coverage via different `revisions`
+// * To observe the failure/repro at LLVM-IR level we need to use `staticlib`
+//   which necessitates `-C prefer-dynamic=false` - without the latter flag,
+//   we would have run into "cannot prefer dynamic linking when performing LTO".
+//
+// The test is restricted to `only-linux`, because the sanitizer-related instrumentation is target
+// specific.  In particular, `___asan_globals_registered` is only used in the
+// `InstrumentGlobalsELF` and `InstrumentGlobalsMachO` code paths.  The `only-linux` filter is
+// narrower than really needed (i.e. narrower than ELF-or-MachO), but this seems ok - having a
+// linux-only regression test should be sufficient here.
+//
+// needs-sanitizer-address
+// only-linux
+//
+// revisions:ASAN ASAN-FAT-LTO
+//[ASAN]          compile-flags: -Zsanitizer=address
+//[ASAN-FAT-LTO]  compile-flags: -Zsanitizer=address -Cprefer-dynamic=false -Clto=fat
+
+#![crate_type="staticlib"]
+
+// The test below mimics `CACHED_POW10` from `library/core/src/num/flt2dec/strategy/grisu.rs` which
+// (because of incorrect handling of `___asan_globals_registered` during LTO) was incorrectly
+// reported as an ODR violation in https://crbug.com/1459233#c1.  Before this bug was fixed,
+// `___asan_globals_registered` would show up as `internal global i64` rather than `common hidden
+// global i64`.  (The test expectations ignore the exact type because on `arm-android` the type
+// is `i32` rather than `i64`.)
+//
+// CHECK: @___asan_globals_registered = common hidden global
+// CHECK: @__start_asan_globals = extern_weak hidden global
+// CHECK: @__stop_asan_globals = extern_weak hidden global
+#[no_mangle]
+pub static CACHED_POW10: [(u64, i16, i16); 4] = [
+    (0xe61acf033d1a45df, -1087, -308),
+    (0xab70fe17c79ac6ca, -1060, -300),
+    (0xff77b1fcbebcdc4f, -1034, -292),
+    (0xbe5691ef416bd60c, -1007, -284),
+];